adaptive blind noise suppression in some speech processing by lindash

VIEWS: 14 PAGES: 11

									                                                www.seminarson.com

                             Abstract

      In many applications of speech processing the noise reveals some
specific features. Although the noise could be quite broadband, there are
a limited number of dominant frequencies, which carry the most of its
energy. This fact implies the usage of narrow-band notch filters that
must be adaptive in order to track the changes in noise characteristics.
In present contribution, a method and a system for noise suppression
are developed. The method uses adaptive notch filters based on second-
order Gray-Markel lattice structure. The main advantages of the
proposed system are that it has very low computational complexity, is
stable in the process of adaptation, and has a short time of adaptation.
Under comparable SNR improvement, the proposed method adjusts only
3 coefficients against 250-450 for the conventional adaptive noise
cancellation systems. A framework for a speech recognition system that
uses the proposed method is suggested.




                                                                        1
                                                   www.seminarson.com




                         1. INTRODUCTION

      The noise existence is inevitable in real applications of speech
processing. It is well known that the additive noise affects negatively the
performance of the speech codecs designed to work with noise-free
speech especially codecs based on linear prediction coefficients (LPC).
Another application strongly influenced by noise is related to the hands
free phones where the background noise reduces the signal to noise ratio
(S/N) and the speech intelligibility.


      Last but not least, is the problem of speech recognition in a noisy
environment. A system that works well in noise-free conditions, usually
shows considerable degradation in performance when background noise
is present It is clear that a strong demand for reliable noise cancellation
methods exists that efficiently separate the noise from speech signal. The
endeavors in designing of such systems can be followed some 20 years
ago    The core of the problem is that in most situations the
characteristics of the noise are not known a priori and moreover they
may change in time. This implies the use of adaptive systems capable of
identifying and tracking the noise characteristics. This is why the
application of adaptive filtering for noise cancellation is widely used.


      The classical systems for noise suppression rely on the usage of
adaptive linear filtering and the application of digital filters with finite
impulse response (FIR). The strong points of this approach are the simple
analysis of the linear systems in the process of adaptation and the
guaranteed stability of FIR structures. It is worth mentioning the
existence of relatively simple and well investigated adaptive algorithms


                                                                           2
                                                  www.seminarson.com

for such kind of systems as least mean squares (LMS) and recursive least
squares (RLS) algorithms. The investigations in the area of noise
cancellation reveal that in some        applications the nonlinear filters
outperform their linear counterparts. That fact is a good motivation for a
shift towards the usage of nonlinear systems in noise reduction .Another
approach    is based on     a   microphone array     instead of   the two
microphones, reference and primary, that are used in the classical noise
cancellation scheme .


      A brief analysis of all mentioned approaches leads to the
conclusion that they try to model the noise path either by a linear or by a
nonlinear    system. Each of these methods has its strengths and
weaknesses. For example, for the classical noise cancellation with two
microphones this is the need of reference signal; for the neural filters -
the fact that as a rule they are slower than classic adaptive filters and
they are efficient only for noise suppression on relatively short data
sequences which is not true for speech processing and finally for
microphone arrays – the need of precise space alignment In present
contribution the approach is slightly different. The basic idea is that in
many applications, for instance, hands-free cellular phones in car
environment howling control in hands-free phones, noise reduction in an
office environment, the noise reveals specific features that can be
exploited. In most instances although the noise might be quite wide-
band, there are always, as a rule, no more than two or three regions of
its frequency spectrum that carry most of the noise energy and the
removal of these dominant frequencies results in a considerable
improvement of S/N ratio. This brings the idea to use notch adaptive
filters capable of tracking the noise characteristics. In this paper a
modification of all-pass structures is used They are recursive, and at the
same time, are stable during the adaptive process. The approach is called
“blind” because there is no need of a reference signal.


                                                                          3
                                                     www.seminarson.com

II. CLASSICAL ADAPTIVE NOISE CANCELLATION


        One of the most wide spread applications of adaptive filtering is
adaptive noise cancellation. Fig. 1 shows the popular scheme for
adaptive noise cancellation using digital FIR filter.




        The basic prerequisite for this realization is the availability of the
two inputs called primary and reference. The primary signal consists of
speech s(n) plus noise n1(n) while the reference signal consists of noise
n(n) alone. The two noise signals n 1(n) and n(n) are correlated and h i(n)
is the impulse response of the noise path from the noise source to the
primary microphone. Assuming that the signals are discretetime and the
sampling period is T=1, the primary input can be written as
        xp(n) = s(n) +n1(n) …………. (1)
where speech signal s(n) and noise signal n 1(n) must be uncorrelated.
        Going through the scheme of Fig. 1 and all mentioned above it is
clear that here noise cancellation is simply the joint process estimation
problem. The system is to reduce the effect of the noise in the primary
input using the correlation between the two noise signals n(n) and n 1(n).
This can be implemented by minimizing the mean-square error E[e2(n)]
where
        e(n) = xp(n) – y(n)……………. (2)
In (2) y(n) is the output signal of the adaptive filter


                           ………….(3)




                                                                             4
                                                   www.seminarson.com

where N is the filter order and w i(n) is the ith coefficient of the adaptive
filter. Having in mind that


                                     …..(4)
the minimization of E[e2(n)] is equivalent to the minimization of the
difference between the signal on the adaptive filter output y(n) and the
noise signal n1(n) present on the primary input. Obviously the better
replica of n1(n) y(n) is, that is, the better the adaptive filter is modeling
the impulse response h i(n) of the noise path, the smaller the difference.
The minimization of E[e2(n)] can be achieved by updating of the adaptive
filter coefficients. Most often the LMS and NLMS algorithms are used, the
latter having the advantage that the step size is relatively independent of
the amplitude of the input signal. According to the scheme in Fig. 1 the
updating equations for LMS and NLMS algorithms are given by




III. ADAPTIVE BLIND NOISE SUPPRESSION (ABNS) SCHEME
       As mentioned in the introduction, the specific features of the noise
in some speech processing applications suggest the usage of narrow-
band notch
filters. They have to meet the following requirements:
      To adapt as fast as possible to the changes in the noise which
       might be quite rapid, for example car engine noise;
      The cancelled portions of the spectrum should be as narrow as
       possible in order to prevent speech signal distortions.
       Both requirements could be met much easier using IIR adaptive
filters instead of FIR adaptive filters. IIR filters are usually avoided



                                                                            5
                                                  www.seminarson.com

because they create a lot of stability problems. To overcome this problem
we use a realization based on second order Gray-Markel lattice circuit
Fig.2. Using this circuit it becomes possible to implement a second order
notch/bandpass section Fig. 3.The advantages of such a realization are
first, it has extremely low pass band sensitivity that means resistance to
quantization effects.   Second, it is very convenient for realization of
adaptive notch filters because it is possible tocontrol independently the
notch frequency and the bandwidth.




Thus if the all-pass function A(z) is




then k1 controls the notch frequency       0   while k2 is related to the
bandwidth
BW via




                                                                         6
                                                   www.seminarson.com

      But, on the other hand, BW is directly connected to the distance
from the pole to the unity-circle. So if we use the structure of Fig. 3 as an
adaptive filter we may fix BW and thus fixing k2 we make constant the
distance from the pole to the unity-circle which means that with this
constraint we obtain     an adaptive IIR filter free of stability problems.
Adapting k1 we may shift the notch frequency around the unity-circle.
Using the basic structure of Fig. 3 and the constraint mentioned above,
the final arrangement of our system
is shown in Fig. 4.
       The system will work in the following manner: each section will
remove one of the dominant frequencies using an appropriate adaptive
algorithm. As shown in Fig. 4 we propose to update only the coefficients
k11, k12,…, k1M, while k2 is a priori determined from equation (9). Thus we
can reduce considerably the number of computations and can guarantee
the stability of the adaptive structure. The number of sections is
determined by the application. Here we introduce the NLMS algorithm for
adjusting the filter coefficients as




where M is the number of sections, ei(n) is the error signal,     is the step
size and   yi„(n) is the derivative of y i(n) with respect to the coefficient
subject of adaptation.




                                                                            7
                                                www.seminarson.com




IV. TEST RESULTS
      The performance of the ABNS method for noise suppression is
assessed by computer simulations. Fig. 5 shows the original speech. The
speech is corrupted with noise from computer cooling fan that is most
often encountered in office environment and the resultant signal is
depicted in Fig. 6. The process of noise suppression is shown in Fig .7.
Here the   system is composed of 3 sections each of them adapting its
coefficient to one of the dominant frequencies in the noise spectrum.



                                                                        8
                                                    www.seminarson.com

Fig.8 presents the trajectories of the filter coefficients. In this experiment
the capability of the system to track the changes in noise signal is tested
as the dominant         frequencies shift from 0.1, 0.2 and 0.4 at the
beginning, to 0.14, 0.23 and       0.36. Here the system does not have
information about the dominant frequencies and adjusts its coefficients
to them, as it works.




                                                                             9
                                                  www.seminarson.com


      Table 1 shows the improvement in SNR as a result of the
application of the proposed system. The obtained results are comparable
to these of the conventional adaptive noise canceller (ANC).The proposed
system is much faster and simpler to implement.




V. A FRAMEWORK FOR A SPEECH RECOGNITION
   SYSTEM USING ABNS
      A block diagram of a speech recognition system is given in Fig. 9. It
consists of the following modules: adaptive blind noise suppression
(ABNS), endpoint detection (EPD), acoustic feature extraction (AFE),
feature normalization (FN) and speech recognition module (SRM).




      Short-time energy and zero-crossing rate are combined to detect
the speech utterance boundaries. Acoustic features of the input speech
are extracted over 20 ms frames. Hamming windows having an overlap of
10 ms are used to calculate Mel Frequency Scale Cepstral Coefficients
and log- energy. Here the speech recognizer can be implemented on the
base of adaptive evolving fuzzy neural networks (EFuNNs) .Since the
input layer of the networks has fixed size, while the segments (words) are
made up of a variable number of frames, a technique for normalization is
needed. A discrete cosine transform (DCT) is applied to the whole
segment, retaining as many parameters as it is necessary. Several




                                                                         10
                                                    www.seminarson.com

application-oriented systems for automatic dialing and robot control are
under development.


                            VI. CONCLUSIONS

       A very efficient adaptive system based on IIR structures for noise
suppression is proposed in this contribution. The main advantages of the
present realization are:
      The adaptive system has a short time of adaptation - about 100
       iterations;
      The system is very simple and flexible, for comparison, here we
       adjust only 3 coefficients against 250-450 for conventional
       adaptive noise cancellation    system;
      The   second-order   lattice   structures   are   stable   during   the
       adaptation that defines the high stability of the whole system.


       The proposed system for noise suppression may be applied in
many situations where the noise reveals the specific features mentioned
in the previous sections and the application of this system could
considerably improve the speech intelligibility.




                                                                            11

								
To top